Wild-Relation-Network

The repo is a PyTorch implementation of Wild Relational Network (WReN) introduced in DeepMind's Measuring abstract reasoning in neural networks (ICML 2018).

Dependencies

Important

• PyTorch (0.4.1)
• TensorBoardX (and Tensorboard)

See requirements.txt for other dependencies.

Usage

Run

python main.py --model <WReN/CNN_MLP/Resnet50_MLP/LSTM> --img_size <input image size> --path <path to your dataset>

Accuracy Plot

The following figure shows the WReN performance we got using the hyper-parameters in the paper.